Anthropic has started letting outside researchers inside its own walls. The company confirmed this week that it has given six employees from METR (Model Evaluation and Threat Research) “employee-like access” to its internal systems, a move that goes well beyond the usual arm’s-length audits that AI labs have relied on until now. The six researchers now carry Anthropic hardware, use Anthropic office space, and tap into internal tools roughly on par with what the company’s own risk-assessment staff can see, according to CNBC.
The announcement lands at a sensitive moment. Anthropic has spent much of 2026 fending off questions about whether its own models can act unpredictably, including warnings about AI self-replication and a separate episode involving a rogue AI system sending a fake police tip. Against that backdrop, handing outside evaluators this level of access to Anthropic AI safety evaluators programs is less a goodwill gesture than an attempt to answer a question regulators and researchers keep asking: who is actually checking the checkers?
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Access Did METR Actually Get
According to Anthropic’s own account, the company has “onboarded six METR employees with company hardware, office space, and access to internal tools and systems that are roughly on par with what our own risk assessment teams have access to,” a description that frames this as something closer to a secondment than a scheduled inspection. The announcement, published on Anthropic’s site and reported by Anthropic, says the METR team’s mandate stretches across model behavior, safety practices, risk management, and the handling of incidents, not just a single model’s output on a test set.
Crucially, Anthropic says it cannot veto what METR concludes about risks, incidents, or internal practices. That is a meaningful departure from the standard NDA-bound audit relationship, where the company being reviewed typically has significant control over what ever becomes public. Anthropic’s only lever is a narrowly defined redaction process, covering customer data, legally privileged material, and model architecture specifics. Even then, METR keeps the right to say publicly that a redaction happened and to argue, without revealing the redacted content itself, whether it believes the redaction was justified.
Why Employee-Like Access Is a Bigger Deal Than It Sounds
Most third-party AI evaluations run on a project basis: a lab ships a model, an outside group gets time-boxed access to test it, and a report comes out weeks or months later. What Anthropic has set up with METR is structurally different because it is ongoing and embedded rather than episodic. Six people showing up with company laptops and badges every day changes the nature of what they can observe, including the day-to-day judgment calls that never make it into a formal model card. That is also precisely why this arrangement draws comparisons to internal audit functions at banks or hospitals, rather than to the external penetration-testing contracts common in cybersecurity.
Three Tiers: How Anthropic Ranks Outside Evaluators
Anthropic says it is developing three tiers of access for outside evaluators, with METR currently sitting at the top tier. That top tier combines the employee-like, embedded access described above with publication rights similar to the middle tier, meaning METR gets both the deepest look inside the company and a strong guarantee that its findings will see daylight. Anthropic has not published a detailed rubric explaining exactly what separates tier one from tier two, so most of what is known about the lower tiers comes from how other evaluators are currently being treated rather than from a formal published framework.
That tiered structure is itself a response to a criticism that has dogged AI safety evaluation for years: that “independent review” can mean almost anything, from a single afternoon demo to a genuinely adversarial audit. By naming tiers at all, Anthropic is inviting outside groups, and its own customers, to ask which tier a given evaluator sits in before taking any safety claim at face value.
SecureBio and Accenture’s Faculty: The Other Watchdogs
METR is not the only outside group working with Anthropic. SecureBio, a biosecurity-focused nonprofit, is separately evaluating Anthropic’s safety work, with a particular focus on the kind of biological and chemical weapons risk that AI labs have flagged as a top concern for frontier models. Accenture’s AI business, branded Faculty, is also working with Anthropic on model evaluations and related assessments, according to the company’s disclosures.
The involvement of a large, blue-chip consultancy like Accenture through Faculty signals something different from METR’s nonprofit, researcher-driven model. Where METR and SecureBio are mission-focused outfits built specifically around AI and biosecurity risk, Faculty brings the kind of enterprise-scale assessment muscle that large corporate and government customers tend to recognize and trust when they are deciding whether to deploy Claude models in regulated environments.
What Anthropic Can and Cannot Redact
The redaction rules are the part of this arrangement most likely to be tested in public. Anthropic’s stated redaction categories are narrow by design: customer data, legally privileged or attorney-client material, and model architecture details that could constitute a competitive or security risk if disclosed. Anything outside those three buckets is, in theory, fair game for METR to describe in its own words.
What makes the arrangement harder for Anthropic to quietly water down is METR’s right to flag a redaction publicly even when it cannot reveal what was redacted. If METR believes Anthropic over-redacted a finding about, say, how a model behaved during an internal incident, it can say so, and argue the point, without needing to disclose the underlying detail. That creates a kind of public pressure valve: Anthropic can still protect sensitive information, but it cannot do so silently.
Why Now: The Pressure Behind the Announcement
Anthropic is not making this move in a vacuum. More than 100 AI experts have signed a public letter arguing that OpenAI and Anthropic specifically need independent safety evaluators with real teeth, not just advisory panels, according to IBTimes. At the same time, both OpenAI and Anthropic have faced regulatory scrutiny over safety failures, including subpoenas from the US Federal Trade Commission, reported by Engadget.
Put those two pressures together and the METR arrangement reads less like a voluntary transparency initiative and more like a pre-emptive move to get ahead of regulators who are already asking hard questions. Anthropic has also been dealing with its own legal exposure this year, including a 1.5 billion dollar copyright settlement that closed out a major piece of litigation. Showing regulators and the public a functioning, embedded, can’t-be-vetoed evaluation process is a far stronger argument than another written safety policy.
Historical Context: From Red-Teaming to Embedded Evaluators
AI safety evaluation has moved through several distinct phases since large language models went mainstream. Early red-teaming exercises, common from 2022 through 2024, were mostly internal: labs hired contractors to probe their own models for a few weeks before a launch, then published a short summary. METR itself emerged during that period as one of the first nonprofits to specialize in formally measuring whether a frontier model could complete tasks associated with catastrophic risk, such as autonomous replication or offensive cyber capability.
SecureBio had already run a pilot review of a previous Anthropic risk report before this expanded arrangement, giving it a track record Anthropic could point to when deciding how much access to grant next. What has changed in 2026 is the depth and permanence of the relationship. Instead of a time-boxed engagement tied to a single model release, METR’s six researchers are now a semi-permanent fixture inside Anthropic, closer in spirit to an internal audit function than to a traditional outside consultancy.
How This Differs From Anthropic’s Earlier Safety Disclosures
Anthropic has published safety-related disclosures before, including its own account of concerning model behavior in the self-replication and fake police tip episodes mentioned above. Those were self-reported. The METR arrangement is different because the reporting entity is not Anthropic itself, and Anthropic has explicitly given up its veto over the conclusions. That is the structural shift worth paying attention to: safety reporting is moving from “here is what we found and chose to tell you” toward “here is what an outside party found, and we cannot stop them from telling you.”
Competitive Comparison: Where Anthropic’s Rivals Stand
Anthropic’s move puts pressure on every other frontier AI lab to explain why it has not done something similar. OpenAI, Anthropic’s closest competitor on both model capability and safety messaging, has not announced an equivalent embedded-access arrangement with an outside evaluator at the scale now described for METR. Given that OpenAI has separately faced its own scrutiny this year, including an employee legal matter involving a recording device, the contrast between a company facing internal security questions and a rival publicly inviting outside staff into its buildings is not a flattering one for OpenAI’s safety optics.
Google DeepMind and Meta both maintain their own internal safety and responsible-AI teams, and both have published model evaluation frameworks, but neither has publicly described handing a nonprofit evaluator full-time, badge-carrying access comparable to what Anthropic has now granted METR. For enterprise buyers comparing Claude’s frontier models against rival offerings, that difference in evaluation transparency is becoming a genuine point of competitive differentiation, not just a public-relations talking point.
Anthropic’s Reported Evaluator Access Tiers
Note that Anthropic has not published a fully detailed public rubric distinguishing Tier 1 from Tier 2 criteria; the above reflects how current evaluators are being described rather than a formally released scoring system.
External AI Safety Evaluators Working With Anthropic
Industry and Market Reaction
Coverage of the announcement has been largely framed around external evaluation as a mechanism to improve safety oversight without necessarily slowing AI development down, rather than as a story with an immediate stock-price or investor angle. That framing matters: Anthropic is privately held, so there is no public market reaction to track the way there would be for a listed company, but the arrangement is being closely watched by enterprise customers and policymakers who care about the credibility of AI safety claims more than quarterly earnings.
Business Insider’s coverage framed the announcement as Anthropic opening its doors to outside safety watchdogs, a characterization that fits the broader narrative Anthropic has been building around “responsible scaling” since co-founder and chief executive Dario Amodei first set out the company’s safety-first positioning. Whether that positioning translates into genuine behavioral change at other labs, rather than just a new round of safety-themed press releases, is the open question analysts are now watching for.
Expert and Industry Voices
Anthropic described the program’s starting point directly: “Today, we’re announcing that we’ve started to give employee-like access to a team from METR, a nonprofit that evaluates frontier AI models,” the company said in its official update, published at anthropic.com.
Anthropic went further in describing the practical shape of that access: “We’ve onboarded six METR employees with company hardware, office space, and access to internal tools and systems that are roughly on par with what our own risk assessment teams have access to,” the company said in the same announcement.
Anthropic also addressed how disputes over redactions would play out in public, noting that METR “is free to characterize our redactions publicly, and say whether it believes the redactions were justified, as long as it does so without revealing the redacted information.” That framing is meant to reassure outside observers that the redaction process is not simply a back door for Anthropic to quietly soften unfavorable findings.
For its part, METR has signaled it intends to be transparent about the limits of its own access, stating that its eventual report “will openly describe the scope, access, and other limitations” of its work, an acknowledgment that even employee-like access has boundaries worth disclosing to readers.
What This Means for Enterprises and Developers Using Claude
For Australian businesses and developers building on Claude models, this arrangement is relevant beyond the headline. Procurement teams evaluating AI vendors for regulated sectors, finance, healthcare, government, increasingly ask for evidence of independent safety review as part of vendor due diligence. Anthropic’s embedded-evaluator model gives those teams something more concrete to point to than a self-published model card, particularly for organizations that have already standardized on Claude over competing chatbot platforms on cost or performance grounds.
It also raises the bar for what “safety-tested” should mean in vendor conversations generally. A developer choosing between frontier model providers now has a genuine, structural difference to weigh: does this vendor merely publish its own safety testing, or does it let an outside group with no obligation to flatter the company walk through its systems and publish what it finds, redactions notwithstanding.
What’s Next: Hugging Face Talks and Further Expansion
Anthropic says discussions with Hugging Face and other organizations about expanding the evaluator program are ongoing, though the company has not detailed the scope or timeline of those talks. Anthropic has also indicated it may expand SecureBio’s access and publication rights once its current biosecurity review wraps up, suggesting the three-tier system is meant to be a ladder evaluators can climb over time rather than a fixed, one-time classification.
None of that is confirmed as locked-in policy yet. Anthropic has not stated publicly that the METR arrangement is permanent, that it will extend automatically to every future model release, or that outside evaluators will gain any formal authority to delay or block a model launch. Those remain open questions that reporters and researchers are likely to keep pressing Anthropic on as the program matures.
Predictions: Where Independent AI Evaluation Goes From Here
- Expect at least one other frontier lab, most likely OpenAI or Google DeepMind, to announce a comparable embedded-evaluator arrangement within the next two to three quarters, driven by competitive pressure rather than regulation alone.
- A public dispute over a specific redaction is likely within the program’s first year, given METR’s explicit right to flag disagreements without revealing redacted content.
- Enterprise and government procurement processes for frontier AI models will increasingly ask vendors to disclose which, if any, evaluator tier they have opened themselves up to, turning this into a genuine sales differentiator.
- Hugging Face’s ongoing talks with Anthropic are likely to formalize into at least a Tier 1 or Tier 2 arrangement, given the momentum already built by the METR and SecureBio relationships.
- Regulatory bodies, including the FTC in the US, will likely cite the METR model as a reference point when drafting future AI safety disclosure requirements for frontier labs.
Related
- Anthropic Warns AI Can Self-Replicate on 15 TV Boxes [2026]
- Anthropic Reveals Rogue AI Sent Fake Police Tip [2026]
- Anthropic’s $1.5B Copyright Settlement Clears Court [2026]
- Claude Opus 5 Launches, Undercuts Fable 5 by 50% [2026]
- OpenAI Engineer Faces 5-Year Max in FBI Bugging Case [2026]
Frequently Asked Questions
What is METR and why does it matter to Anthropic?
METR, short for Model Evaluation and Threat Research, is a nonprofit that evaluates frontier AI models for catastrophic risk. It matters to Anthropic because the company has now given six METR employees employee-like access to its internal systems, making METR one of the most deeply embedded outside evaluators working with any major AI lab.
How many METR employees have access to Anthropic’s systems?
Six METR employees have been onboarded with company hardware, office space, and access to internal tools and systems that Anthropic says are roughly comparable to what its own risk-assessment teams can access.
Can Anthropic block METR’s findings from being published?
No. Anthropic says it cannot veto METR’s conclusions about risks, incidents, or safety practices. It can only require narrowly defined redactions covering customer data, legally privileged information, and model architecture, and METR can still publicly flag that a redaction occurred.
What is SecureBio’s role in Anthropic’s safety evaluations?
SecureBio is a biosecurity-focused nonprofit evaluating Anthropic’s safety work, with a focus on biological and chemical weapons risk. It is currently reviewing the relevant chapter of Anthropic’s August risk report, following an earlier pilot review of a previous report.
Is Anthropic working with Hugging Face too?
Anthropic has said discussions with Hugging Face and other organizations are ongoing, but the scope and status of any arrangement with Hugging Face have not been fully specified publicly.
What are Anthropic’s three tiers of evaluator access?
Anthropic says it is developing three tiers of outside evaluator access, ranging from narrowly scoped reviews that need company approval before publishing, up to METR’s current Tier 3, which combines employee-like embedded access with strong publication rights.
Does this access apply to all of Anthropic’s AI models?
Anthropic has not publicly confirmed that METR’s access automatically extends to every future model release. The current arrangement covers the model behavior, safety practices, risk management, and incident handling that METR’s six embedded staff can observe today.
Why is independent AI safety evaluation becoming a bigger issue in 2026?
Pressure has built from multiple directions in 2026, including a public letter signed by more than 100 AI experts calling for stronger independent evaluators, regulatory scrutiny such as FTC subpoenas over safety failures, and a string of publicly disclosed incidents involving frontier AI models behaving unpredictably.